Mercury 3 vs GPT-5.6 Sol: 20 Times Cheaper per Token, and the Tokens Are the Whole Argument

The interesting thing about putting Mercury 3 next to GPT-5.6 Sol is that Sol is already the cheap end of O​penAI's frontier and Mercury 3 is cheaper again by a margin that makes the word "cheap" do no work. Inception Labs released Mercury 3 on October 19, 2026 at $0.20 per million input tokens and $0.70 per million output, halved to $0.10 and $0.35 during a launch window with no stated end date. GPT-5.6 Sol, which O​penAI shipped on July 9, 2026, lists at $4.00 and $20.00 below 272,000 tokens and $8.00 and $30.00 above it. On output that is a 57-fold gap at Mercury's launch price and a 28-fold gap at list — and unlike most such gaps, it does not shrink when you measure it per answer, because Sol's advantage in the earlier G​PT-6 comparisons was token frugality and Sol is not frugal in the way GPT-6 Sol was. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

Both models will take a million tokens of input in one call and both emit text only in return. Where they part is what happens around the tokens: Mercury 3 is a diffusion language model that commits blocks of positions at once and runs at a vendor-measured 1,068 tokens per second, while GPT-5.6 Sol is an autoregressive frontier model that accepts images and files as well as text, and which Artificial Analysis measured at 87.7 tokens per second for $1.99 per Intelligence Index task. The comparison is therefore less about which model is better, which nobody has measured for Mercury, and more about what the same workload costs on each once retries and wasted tokens are counted.

Two rate cards, read carefully

Sol's card has a boundary in it that Mercury's does not, and the boundary is the most consequential structural difference between the two products. Everything up to 272,000 input tokens bills at $4.00 and $20.00. Anything above that reprices the entire request at $8.00 and $30.00 — not the excess tokens, the whole thing. A 300,000-token prompt costs $2.40 of input on Sol, which is more than a 250,000-token prompt would cost at $4.00 by a factor of more than two, for 20% more text.

• Price — Mercury 3: $0.20 / $0.70 per million, $0.10 / $0.35 at launch, cached input 90% off the input rate. GPT-5.6 Sol: $4.00 / $20.00 per million under 272,000 tokens, $8.00 / $30.00 above, cached input at $0.40 with a $5.00 per-million write cost.

• Context and output — Mercury 3: 1,000,000 tokens in, 131,072 out. GPT-5.6 Sol: 1,050,000 in, 128,000 out.

• Modalities — Mercury 3: text only. GPT-5.6 Sol: text, image and file in, text out.

• Speed — Mercury 3: 1,068 tokens per second, Inception-measured. GPT-5.6 Sol: 87.7 tokens per second on Artificial Analysis's harness, with a time to first token of 91.8 seconds on its default long-prompt workload.

• Independent scores — Mercury 3: none published; no Artificial Analysis page exists. GPT-5.6 Sol: 47 on the Intelligence Index, 77.4 on the coding index, 94.1 on GPQA Diamond, 49.5 on Humanity's Last Exam, 84.0 on long-context recall.

• Release — Mercury 3: October 19, 2026. GPT-5.6 Sol: July 9, 2026.

One number on Sol's card deserves its own sentence, because it is the number a latency budget is made of. Artificial Analysis records a 91.8-second median time to first token for GPT-5.6 Sol on the configuration it tests. That is not a typo and it is not a cold start; it is what a reasoning-first model does when it thinks before it speaks on a long prompt. Mercury 3's launch coverage includes third-party latency reports — Octen measured a median generation time of 0.9 seconds and a median end-to-end response of 2.2 seconds on its workload — but those are customer testimonials in a vendor announcement, not a neutral harness, and different workloads. The direction is not in doubt; the magnitude is unresolvable without a common test.

Why token count decides this one, and how to check it for your workload

The per-token gap is 28-fold at list and 57-fold at launch, so the only way Sol closes it is by producing far fewer tokens for the same work. That is exactly what happened in the GPT-6 Sol comparison, where Sol spent about 31,000 output tokens per index task against Claude Sonnet 5.5's 193,000 and ended up cheaper per task despite a 20-fold higher output rate. It does not obviously happen here: GPT-5.6 Sol costs $1.99 per Intelligence Index task, where GPT-6 Sol cost about $1.05 on the same harness, so Sol is roughly twice as expensive per answer as its successor and around 40 times as expensive per answer as a $0.05 Mercury 3 task would be if Mercury consumed the same token volume.

That "if" is doing all the work, and it is worth stating as a test rather than a claim. Take a hundred of your own prompts, run them through Sol, and record output tokens per prompt and whether the answer needed a retry. Do the same on Mercury 3. Multiply by the rate cards. The result is your answer, and it takes an afternoon rather than a benchmark argument.

Three shapes of result are common enough to predict in advance.

• Short, high-volume text work — classification, extraction, routing, short summaries, 150 output tokens a call. Sol's token frugality has nothing to act on, so the rate card is the whole story: 28 to 57 times cheaper on Mercury 3, before the discount.

• Reasoning-heavy work — a multi-step analysis that Sol spends 20,000 reasoning tokens on. If Mercury 3 needs two attempts to reach the same answer, the effective gap drops from 28-fold to 14-fold; if it needs five, it drops to under six-fold. Still a large saving, no longer a rounding error in the decision.

• Anything involving an image — a screenshot, a scanned page, a chart. Mercury 3 returns an error. The comparison does not exist, and the price of Sol's multimodal input is whatever Sol charges.

The discount, and why the honest cost model uses list

Mercury 3's $0.35 output rate is a promotion. Inception has published the list price and the discount but not the end date, which means a cost model built on $0.35 has a liability in it that nobody can size. On the short-answer workload above, the difference between launch and list pricing is 2x — the same 2x that separates a good quarter from a bad one, arriving without warning.

Sol's card has the opposite property. It is stable, it has a documented boundary, and it has a documented cache structure: $0.40 per million on reads, a $5.00 per-million write cost, which is 25% above the input rate. For a workload that resends the same 30,000-token preamble on every step of an agent loop, that write premium is worth calculating before dismissing — Sol pays 25% more once and 90% less on every subsequent read, and at ten or more steps per task the amortised input cost falls below $1.00 per million.

Mercury 3's cached rate is $0.02 per million at list, $0.01 during the launch window, with no published write premium. That is the single strongest number in its favour for exactly that workload: a twenty-step agent loop with a 30,000-token preamble costs $0.60 of cached input on Mercury at list, against roughly $0.40 per million's worth of reads on Sol after the write — comparable in absolute terms, and the comparison stays comparable only because Sol's caching is unusually good. Strip the caching from both and the gap is 20-fold again.

What the routing layer does with a 20-fold gap and no neutral score

GPT-5.6 Sol is on the OrcaRouter catalogue at O​penAI's own rates — $4.00 and $20.00 per million, $8.00 and $30.00 above the 272,000-token boundary, $0.40 cached reads — behind the same OpenAI-compatible endpoint as everything else we route, with no markup added on top. Mercury 3 is not on the catalogue; our model API returns nothing for any Inception model, so anyone calling it goes to Inception's own platform, where new accounts get 100 million free tokens, or to one of the several third-party platforms that host it.

The gap is large enough that the tempting move is to switch the whole workload and see what breaks. The case for not doing that is not caution for its own sake; it is that Mercury 3 has one quality result, produced inside Inception's own agent harness on 50 live Mind2Web tasks, where it completed 50% and reportedly did so at the same token volume and step count as a frontier model. That is a genuinely interesting result — live sites, not a snapshot — and it is also unfalsifiable from outside, because the harness and the scoring belong to the seller.

What a routing rule buys in that situation is a bounded experiment. Send a defined slice of traffic to Mercury 3, keep Sol behind it as the fallback path, and let automatic failover cover the case where the new model times out or errors. The slice is a number in a config file, not a code path; widening it from 10% to 100%, or narrowing it to zero when the discount ends, is a change to that number.

There is also a third option that neither vendor's page will tell you about, and it is the one most teams land on once they run the numbers: GPT-6 Sol. It carries the same $4.00 and $20.00 rate card as GPT-5.6 Sol, sits one tier below the G​PT-6 flagship, and costs about $1.05 per Intelligence Index task against Sol's $1.99. If the reason you are looking at Mercury 3 is that Sol is too expensive per answer, the cheaper answer to that problem may be a newer model at the same price rather than an unmeasured one at a twentieth of it.

Where this lands

For text-only, high-volume work, the arithmetic is not close and does not depend on anything unverified: at list pricing Mercury 3 is 20 times cheaper on input and 28 times cheaper on output than GPT-5.6 Sol, with a context window of the same size and no boundary to cross. If your workload is a million short calls a month, that is the decision, and the risk is measurable within a day of traffic.

For work that reasons for twenty thousand tokens, reads a screenshot, or runs unattended for hours with tools in the loop, Sol is the model with a documented behaviour and a stable price, and the 91.8-second time to first token is a real number you can plan around rather than a claim you have to test. That is not a small thing to be paying for.

The one measurement that would end the argument is Mercury 3's cost per Intelligence Index task on the harness that produced Sol's $1.99. It does not exist, no date has been given for it, and until it does every per-answer figure in this article for Mercury 3 is an assumption with a rate card attached. The rate card, at least, is real — and at 20-fold, it is wide enough to be worth testing before it is worth believing.